consult

Cross-tool AI consultation: query another AI CLI tool and return the response.

When to Use

Invoke this skill when:

• User wants a second opinion from a different AI tool
• User asks to consult, ask, or cross-check with gemini/codex/claude/opencode/copilot
• User needs to compare responses across AI tools
• User wants to validate a decision with an external AI

Arguments

Parse from

$ARGUMENTS

--tool

--effort

--model

--context

--continue

Question text is everything in

$ARGUMENTS

Provider Configurations

Claude

Command: env -u CLAUDECODE claude -p "QUESTION" --output-format json --model "MODEL" --max-turns TURNS --allowedTools "Read,Glob,Grep"
Session resume: --resume "SESSION_ID"

Models: claude-haiku-4-5, claude-sonnet-4-6, claude-opus-4-6

Parse output:

JSON.parse(stdout).result

JSON.parse(stdout).session_id

npx -y @anthropic-ai/claude-code-acp

Gemini

Command: gemini -p "QUESTION" --output-format json -m "MODEL"
Session resume: --resume "SESSION_ID"

Models: gemini-2.5-flash, gemini-2.5-pro, gemini-3-flash-preview, gemini-3-pro-preview, gemini-3.1-pro-preview

Parse output:

JSON.parse(stdout).response

JSON.parse(stdout).session_id

--resume

gemini

Codex

Command: codex exec "QUESTION" --json -m "MODEL" {SKIP_GIT_FLAG} -c model_reasoning_effort="LEVEL"
Session resume: codex exec resume "SESSION_ID" "QUESTION" --json -m "MODEL" {SKIP_GIT_FLAG} -c model_reasoning_effort="LEVEL"
Session resume (latest): codex exec resume --last "QUESTION" --json -m "MODEL" {SKIP_GIT_FLAG} -c model_reasoning_effort="LEVEL"

Note:

codex exec

-q

codex

{SKIP_GIT_FLAG}

• inside trusted git repo: empty string
• trusted non-repo execution:

  --skip-git-repo-check

Models: gpt-5.3-codex

Parse output:

JSON.parse(stdout).message

codex resume SESSION_ID

JSON.parse(stdout).session_id

~/.codex/sessions/

codex exec resume "SESSION_ID" "follow-up prompt" --json -m "MODEL" {SKIP_GIT_FLAG} -c model_reasoning_effort="LEVEL"

--last

npx -y @zed-industries/codex-acp

OpenCode

Command: opencode run "QUESTION" --format json --model "MODEL" --variant "VARIANT"
Session resume: opencode run "QUESTION" --format json --model "MODEL" --variant "VARIANT" --continue (most recent) or --session "SESSION_ID"
With thinking: add --thinking flag

Models: 75+ via providers (format:

provider/model

opencode/

github-copilot/

amazon-bedrock/

google/

github-copilot/gemini-3.1-pro-preview

opencode/big-pickle

amazon-bedrock/anthropic.claude-opus-4-6-v1

opencode models

Free models:

opencode/big-pickle

opencode/gpt-5-nano

opencode/minimax-m2.5-free

opencode/trinity-large-preview-free

Parse output: OpenCode outputs newline-delimited JSON events. Each line is a JSON object with a

type

type === "text"

part.text

part.content

part.text

type: "text"

step_start

tool_use

text

step_finish

sessionID

sessionID

Session ID: Available in every event as

event.sessionID

ses_xxxxx

--session SESSION_ID

--continue

--session

--session SESSION_ID

--continue

opencode acp

Copilot

Command: copilot -p "QUESTION"

Models: claude-sonnet-4-6 (default), claude-opus-4-6, claude-haiku-4-5, gpt-5

Parse output: Raw text from stdout Continuable: No ACP adapter:

copilot --acp --stdio

Kiro

ACP-only provider. No CLI mode for external consultation.
Command: node acp/run.js --provider="kiro" --question-file="{AI_STATE_DIR}/consult/question.tmp" --timeout=120000

Kiro is available only via ACP transport. It requires

kiro-cli

Parse output: Via ACP runner (

JSON.parse(stdout)

kiro-cli acp

Input Validation

Before building commands, validate all user-provided arguments:

• --tool: MUST be one of: gemini, codex, claude, opencode, copilot, kiro. Reject all other values.
• --effort: MUST be one of: low, medium, high, max. Default to medium.
• --model: Allow any string, but quote it in the command.
• --continue=SESSION_ID: If provided, SESSION_ID MUST match

  ^(?!-)[A-Za-z0-9._:-]+$

  . Reject values that contain spaces, leading dashes, or shell metacharacters.
• --context=file=PATH: MUST resolve within the project directory. Reject absolute paths outside cwd. Additional checks:
  1. Block UNC paths (Windows): Reject paths starting with

     \\

     or

     //

     (network shares)
  2. Resolve canonical path: Use the Read tool to read the file (do NOT use shell commands). Before reading, resolve the path: join

     cwd + PATH

     , then normalize (collapse

     .

     ,

     ..

     , resolve symlinks)
  3. Verify containment: The resolved canonical path MUST start with the current working directory. If it escapes (via

     ..

     , symlinks, or junction points), reject with:

     [ERROR] Path escapes project directory: {PATH}

  4. No shell access: Read file content using the Read tool only. Never pass user-provided paths to shell commands (prevents injection via path values)

Command Building

Given the parsed arguments, build the complete CLI command. All user-provided values MUST be quoted in the shell command to prevent injection.

Step 1: Resolve Model

If

--model

Step 1b: Trust Gate for Codex

--skip-git-repo-check

Before using any Codex template, resolve

{SKIP_GIT_FLAG}

1. Verify the consultation is running from the current project working directory (the same workspace where

   /consult

   was invoked), not an arbitrary external path.
2. Verify the resolved active tool is Codex (flag, NLP, picker, or restored

   --continue

   session).
3. Run

   git rev-parse --is-inside-work-tree

   in the current working directory:
   • if true: set

     {SKIP_GIT_FLAG}

     to empty string
   • if false and checks 1-2 passed: set

     {SKIP_GIT_FLAG}

     to

     --skip-git-repo-check

4. If checks 1-2 fail, reject execution with

   [ERROR] Refusing Codex --skip-git-repo-check outside trusted working directory

   .

Codex templates in this skill assume this trust gate has already passed.

Step 2: Build Command String

Use the command template from the provider's configuration section. Substitute QUESTION, MODEL, TURNS, LEVEL, VARIANT, and SKIP_GIT_FLAG with resolved literal values.

{SKIP_GIT_FLAG}

SKIP_GIT_FLAG

If continuing a session:

• Claude or Gemini: append

  --resume "SESSION_ID"

  to the command.
• Codex: use

  codex exec resume "SESSION_ID" "QUESTION" --json -m "MODEL" {SKIP_GIT_FLAG} -c model_reasoning_effort="LEVEL"

  instead of the standard command. Use

  --last

  instead of a session ID for the most recent session.
• OpenCode: append

  --session SESSION_ID

  to the command. If no session_id is saved, use

  --continue

  instead (resumes most recent session). If OpenCode at max effort: append

  --thinking

  .

Step 3: Context Packaging

If

--context=diff

git diff 2>/dev/null

--context=file=PATH

Step 4: Safe Question Passing

User-provided question text MUST NOT be interpolated into shell command strings. Shell escaping is insufficient --

$()

Required approach -- pass question via stdin or temp file:

1. Write the question to a temporary file using the Write tool (e.g.,

   {AI_STATE_DIR}/consult/question.tmp

   )

   Platform state directory:

   • Claude Code:

     .claude/

   • OpenCode:

     .opencode/

   • Codex CLI:

     .codex/

2. Build the command using the temp file as input instead of inline text:

env -u CLAUDECODE claude -p - --output-format json --model "MODEL" --max-turns TURNS --allowedTools "Read,Glob,Grep" < "{AI_STATE_DIR}/consult/question.tmp"

env -u CLAUDECODE claude -p - --output-format json --model "MODEL" --max-turns TURNS --allowedTools "Read,Glob,Grep" --resume "SESSION_ID" < "{AI_STATE_DIR}/consult/question.tmp"

gemini -p - --output-format json -m "MODEL" < "{AI_STATE_DIR}/consult/question.tmp"

gemini -p - --output-format json -m "MODEL" --resume "SESSION_ID" < "{AI_STATE_DIR}/consult/question.tmp"

codex exec "$(cat "{AI_STATE_DIR}/consult/question.tmp")" --json -m "MODEL" {SKIP_GIT_FLAG} -c model_reasoning_effort="LEVEL"

codex exec resume "SESSION_ID" "$(cat "{AI_STATE_DIR}/consult/question.tmp")" --json -m "MODEL" {SKIP_GIT_FLAG} -c model_reasoning_effort="LEVEL"

codex exec resume --last "$(cat "{AI_STATE_DIR}/consult/question.tmp")" --json -m "MODEL" {SKIP_GIT_FLAG} -c model_reasoning_effort="LEVEL"

opencode run - --format json --model "MODEL" --variant "VARIANT" < "{AI_STATE_DIR}/consult/question.tmp"

opencode run - --format json --model "MODEL" --variant "VARIANT" --session "SESSION_ID" < "{AI_STATE_DIR}/consult/question.tmp"

opencode run - --format json --model "MODEL" --variant "VARIANT" --continue < "{AI_STATE_DIR}/consult/question.tmp"

copilot -p - < "{AI_STATE_DIR}/consult/question.tmp"

3. Delete the temp file after the command completes (success or failure). Always clean up to prevent accumulation.

Model and session ID values are controlled strings (from pickers or saved state) and safe to quote directly in the command. Only the question contains arbitrary user text and requires the temp file approach. The temp file path (

{AI_STATE_DIR}/consult/question.tmp

Provider Detection

Cross-platform tool detection:

• Windows:

  where.exe TOOL 2>nul

  -- returns 0 if found
• Unix:

  which TOOL 2>/dev/null

  -- returns 0 if found

Check each tool (claude, gemini, codex, opencode, copilot, kiro) and return only the available ones.

ACP Transport

ACP (Agent Client Protocol) is an alternative transport to CLI subprocess invocation. When available, ACP provides structured JSON-RPC 2.0 communication, session persistence, and streaming responses via a universal protocol supported by all major AI coding tools.

ACP Provider Adapters

npx -y @anthropic-ai/claude-code-acp

gemini

npx -y @zed-industries/codex-acp

copilot --acp --stdio

kiro-cli acp

opencode acp

Transport Selection

1. Check ACP availability for the target provider (see ACP Detection below)
2. If ACP available: use ACP transport (preferred - standardized protocol, session persistence)
3. If ACP unavailable: fall back to CLI transport (existing behavior above)

The output envelope is identical regardless of transport. Downstream consumers (session management, debate orchestrator, output parsing) are transport-agnostic.

ACP Command Template

All ACP providers use the same command pattern via the ACP runner script:

node acp/run.js --provider="PROVIDER" --question-file="{AI_STATE_DIR}/consult/question.tmp" --timeout=TIMEOUT_MS [--model="MODEL"] [--session-id="SESSION_ID"]

node acp/run.js --provider="claude" --question-file="{AI_STATE_DIR}/consult/question.tmp" --timeout=120000 --model="MODEL"

node acp/run.js --provider="gemini" --question-file="{AI_STATE_DIR}/consult/question.tmp" --timeout=120000 --model="MODEL"

node acp/run.js --provider="codex" --question-file="{AI_STATE_DIR}/consult/question.tmp" --timeout=120000 --model="MODEL"

node acp/run.js --provider="opencode" --question-file="{AI_STATE_DIR}/consult/question.tmp" --timeout=120000 --model="MODEL"

node acp/run.js --provider="copilot" --question-file="{AI_STATE_DIR}/consult/question.tmp" --timeout=120000

node acp/run.js --provider="kiro" --question-file="{AI_STATE_DIR}/consult/question.tmp" --timeout=120000

Parse output: Same as CLI transport -

JSON.parse(stdout)

JSON.parse(stdout).session_id

--session-id="SESSION_ID"

ACP Detection

Run ACP detection alongside CLI detection. For each provider:

node acp/run.js --detect --provider="PROVIDER"

Returns on success (exit 0):

{"provider": "claude", "acp_available": true, "name": "Claude"}

Returns on failure (exit 1):

{"provider": "claude", "acp_available": false, "name": "Claude", "reason": "npx not found on PATH"}

Kiro note: Kiro is ACP-only - it has no CLI mode for external consultation. It only appears as available when

kiro-cli

Session Management

Save Session

After successful consultation, save to

{AI_STATE_DIR}/consult/last-session.json

{
  "tool": "claude",
  "model": "opus",
  "effort": "high",
  "session_id": "abc-123-def-456",
  "timestamp": "2026-02-10T12:00:00Z",
  "question": "original question text",
  "continuable": true,
  "transport": "acp"
}

The

transport

"acp"

"cli"

--continue

"cli"

AI_STATE_DIR

• Claude Code:

  .claude/

• OpenCode:

  .opencode/

• Codex CLI:

  .codex/

Load Session

For

--continue

• tool (from saved state)
• session_id (for --resume flag)
• model (reuse same model)

Before using restored values, re-validate them:

• tool must still be in allow-list: gemini, codex, claude, opencode, copilot, kiro
• session_id must match

  ^(?!-)[A-Za-z0-9._:-]+$

• model must match

  ^[A-Za-z0-9._:/-]+$

  (reject spaces and shell metacharacters)
• if either check fails, reject with

  [ERROR] Invalid restored session data

  and do not build a command

If session file not found, warn and proceed as fresh consultation.

Output Sanitization

Before including any consulted tool's response in the output, scan the response text and redact matches for these patterns:

sk-[a-zA-Z0-9_-]{20,}

[REDACTED_API_KEY]

sk-proj-[a-zA-Z0-9_-]{20,}

[REDACTED_API_KEY]

sk-ant-[a-zA-Z0-9_-]{20,}

[REDACTED_API_KEY]

AIza[a-zA-Z0-9_-]{30,}

[REDACTED_API_KEY]

ghp_[a-zA-Z0-9]{36,}

[REDACTED_TOKEN]

gho_[a-zA-Z0-9]{36,}

[REDACTED_TOKEN]

github_pat_[a-zA-Z0-9_]{20,}

[REDACTED_TOKEN]

ANTHROPIC_API_KEY=[^\s]+

ANTHROPIC_API_KEY=[REDACTED]

OPENAI_API_KEY=[^\s]+

OPENAI_API_KEY=[REDACTED]

GOOGLE_API_KEY=[^\s]+

GOOGLE_API_KEY=[REDACTED]

GEMINI_API_KEY=[^\s]+

GEMINI_API_KEY=[REDACTED]

AKIA[A-Z0-9]{16}

[REDACTED_AWS_KEY]

ASIA[A-Z0-9]{16}

[REDACTED_AWS_KEY]

Bearer [a-zA-Z0-9_-]{20,}

Bearer [REDACTED]

Apply redaction to the full response text before inserting into the result JSON. If any redaction occurs, append a note:

[WARN] Sensitive tokens were redacted from the response.

Output Format

Return a plain JSON object to stdout (no markers or wrappers):

{
  "tool": "gemini",
  "model": "gemini-3.1-pro-preview",
  "effort": "high",
  "duration_ms": 12300,
  "response": "The AI's response text here...",
  "session_id": "abc-123",
  "continuable": true
}

Install Instructions

When a tool is not found, return these install commands:

npm install -g @anthropic-ai/claude-code

npm install -g @openai/codex

npm install -g opencode-ai

brew install anomalyco/tap/opencode

gh extension install github/copilot-cli

Error Handling

"response": "Timeout after 120s"

"response": "No output received"

Integration

This skill is invoked by:

• consult-agent

  for

  /consult

  command
• Direct invocation:

  Skill('consult', '"question" --tool=gemini --effort=high')

Example:

Skill('consult', '"Is this approach correct?" --tool=gemini --effort=high --model=gemini-3.1-pro-preview')